AI is about to be challenged for copyright

[idraw]

AI is about to be challenged for copyright 

Big names are taking lawsuits to court about it, may set precedents as to usage, pertaining specifically to art, especially if you are using AI to create products on PODs. Getty is one of the big names mentioned.
<https://arstechnica.com/tech-policy/2023/04/stable-diffusion-copyright-lawsuits-could-be-a-legal-ear...>
Those using AI to create art for sale may want to hold off till this rolls thru the courts with a decision about copyright. (Unless you have deep pockets and a dynamite legal team)

[Barbara]

If lawsuits continue, and we can be sure they will, what happens if an artist or photographer recognizes their work within, say, a poster here on Zazzle? It can certainly happen. At the very least, they'll issue a take-down notice if not a lawsuit.

AI is very alluring, like a siren's call, and it can only become more and more popular, much to the detriment of real art. Troubled waters ahead.

Colorwash's Home

[KeegansCreation]

An artist or photographer will not recognize their work within any Zazzle designs because it is far too transformative. It does not collage (so there are no recognizable fragments). It takes the concept of the scraped image as defined by how the image was tagged by whoever uploaded it and makes new images based on concepts. 

Overfitting of copyright protected images is rare but possible. If an image is in many, many places  across the internet it will get scraped many, many times and there will be an enormous number of copies in the dataset. This will be diluted to meaninglessness in most cases because a non-unique theme will have many unique copyrighted images with the same tags. But for images which don't have a lot of competition for the tags and are also spread far and wide across the internet, i.e. is famous, you can generate something which is dangerously close to the original  copyrighted image. This is called overfitting.

I belong to an AI facebook group and people were discussing overfitting with the Afghan Girl, that famous photo of a beautiful and staring Afghan girl. Her eyes are famous by now because of that photo. The thing is, there are not a lot of photos of Afghan girls on the internet. There are some (per Google Images) but not enough to compete with the sheer ubiquity of that famous photo. Goodness knows how many iterations there are of it in the dataset Stability used. That is overfitting.

As an experiment, I went to Midjourney and typed in "photo of an Afghan girl". I got hit with a warning that this was a forbidden phrase and further attempts to generate such an image would get me banned, as would attempts to reword the phrase (so I didn't go typing "girl in Afghanistan"). So that's a safeguard. The other safeguard is that any image famous enough to be subject to overfitting is also going to be caught by Zazzle's bots. 

The people bringing the suit don't claim that images generated violate copyright. They claim that training per se violates copyright regardless of unique output. That is an untested claim and it may take years to be sorted out in court. But the output images are not under threat, just the companies that did the training.

[KeegansCreation]

I want to add in another thing that is adjacent to overfitting, the names of living artists. Artist's names are tags. Artists with lots of images on the internet tagged with their name could have their style imitated by the AI. Style is not copyrightable but still there is some legal thin ice there. This was brought up by concept artist Greg Rutkowski. When he googled his name, it brought up not just his own art but images that he never made but yet were tagged with his name and done in a style similar to his. The images didn't violate copyright because he never made them (although they were something he plausibly could make in the future in that style) so he didn't go after the people who uploaded them. But he did address the AI companies that neither his name nor the name of any living artist should be used as an image prompt. Style can't be copyrighted but there is an argument that names are trademarks.

Shortly after, people in the AI facebook group reported that the names of living artists had become null prompts. If they asked for "X painted in the style of Greg Rutkowski" (or other living artists), nothing happened. They didn't get a ban warning but they also didn't get an image that looked like it could plausibly have been done by him. Midjourney has released a new version which is far stricter about prompts (possibly because of this lawsuit) so I'm not about to go experimenting with names of living artists lest I get banned. 

[KeegansCreation]

If Stable Diffusion, DALL E2 and Midjourney go kerflooey, I'll hang my hat on Adobe Firefly. It's currently in beta and I have access to the beta (any Adobe subscriber can request access) although no commercial use until it's out of beta (hopefully by October's annual Photoshop new release). They trumpet quite loudly how they trained only on public domain, licensed content, and Adobe Stock. Nevertheless, despite this seemingly small dataset, it's quite good.

Midjourney is a bit dicey since they're named in the Getty lawsuit since they did scrape Getty Images for training. A lot hangs in the balance of that lawsuit, specifically, is scraping permissable for training even if copyright protected? The output doesn't violate copyright unless you try very, very hard to make it do so. (I tried as an experiment and got hit with a warning from Midjourney that further attempts would get me banned). The big question is if training (rather than output) violates copyright. I am not worried about images I have already made, they don't violate copyright. But Midjourney could get hit with enough of a fine to put it out of business. We will see. It's a pretty huge question and has major implications. As the article notes, it will mean AI be more expensive and licensing will be a hurdle for startups. It will be mainly in the hands of big companies with their own big datasets which are already theirs, such as Adobe.

DALL-E2 is on sturdier ground since it partnered with Shutterstock (and now Shutterstock has an AI of their own which is probably a version of DALL-E2 per the partnership). Nevertheless they might get sued later if they trained on copyrighted material (that is not Shutterstock).

It will be interesting to see how this all shakes out. Adobe will certainly come out on top (unless it turns out they were lying about not training on unlicensed copyrighted material) so Firefly will be a safe bet no matter how this turns out.

[almdrs]

Now I know why watermarks were not recommended.

[Barbara]

@KeegansCreation: Everything you said in your posts indicates that, in a way, lawsuits will make AI better. My one worry at this point is that people might not support human artists as they once did. There are still photographers who love film and even a few who do Daguerreotypes, so why not artists who use paint on canvas and paper and who attract an audience and buyers?

Colorwash's Home

[KeegansCreation]

My prediction is that human artists who work in the physical world (non-digital) will be fine, perhaps elevated. The people who were going to pay 300$ for a real watercolor at a seaside tourist town art fair still will. That they could get a simulation of a watercolor from an AI won't stop them since they can already get a simulation of a watercolor right here on Zazzle and many other websites. There is value in the physical.

My prediction is that the artists who take the worst initial hit will be the photographers who sell to stock sites. Once AI is more controllable, Shutterstock and Adobe Stock (which currently have AI) will have no need for more photos. Currently there is an element of randomness since "reaching in" to latent space is still a black box, but it will become more controllable over time.

Next I think the market for commercial animators and game artists will tighten up. Fewer people can make more content using AI tools. Or perhaps the number of these artists will stay the same but the job itself will change to be more of a technician and less of an artist. On a reddit thread I came across, a game artist was sad because his employer had embraced AI and his job had morphed from creating game assets using his own creativity to editing and fixing the mistakes of the assets the AI made. He felt he'd been turned into a technician rather than an artist although he still had the same job. That sort of thing will become more common.

[Connie]

Are Shutterstock and Adobe requiring that AI-generated "stock photos" are marked AI? I should think there are still people who want ACTUAL PHOTOS, not computer generated fake images.

[idraw]

If you are thinking about using AI on Zazzle, you might want to read what I just posted in this link:
<https://community.zazzle.com/t5/general-zazzle-discussion/how-to-post-and-tag-ai-generated-products-...>
Those already posting AI  inclusive material on Zazzle especially need to read it.

[CrazyMermaid]

Thanks for all this research. It's a big help. 

[idraw]

Look at the list of signatories,
<https://futureoflife.org/open-letter/pause-giant-ai-experiments/>